KR20230060439A - Method and system for detecting recaptured image method thereof - Google Patents

Abstract

The present disclosure relates to a method for identifying whether or not a real-life photographed image is present and a system to which the method is applied. A method for identifying whether a real-life photographed image according to the present disclosure includes the steps of acquiring a first model learned to determine whether a whole image is a real-photographed image and a second model learned to determine whether a partial image is a real-photographed image; Steps of inputting image data to the first model, obtaining first output data from the first model, obtaining second output data from the second model, and the first output data and the second output data. The method may include outputting data on whether the target image is a real-life photographed image by using all of the output data.

Description

Method and system for identifying actual photographed images {METHOD AND SYSTEM FOR DETECTING RECAPTURED IMAGE METHOD THEREOF}

The present disclosure relates to a method for identifying whether or not a real-life photographed image is present and a system to which the method is applied. More specifically, it relates to a method of identifying whether or not a real-life image is captured by comprehensively analyzing output data of a plurality of artificial intelligence models, and a system to which the method is applied.

Recently, non-face-to-face authentication methods have been widely used for purposes including financial transactions. However, in the conventional method, using the fact that the person requesting authentication cannot be directly identified, there is a problem in that a third party steals another person's ID card image and attempts identity authentication.

In addition, in the related art, a three-dimensional environment confirmation method requiring a user to take a photograph taken at a different angle or a physical additional confirmation method requiring an additional photograph of a user's face have been used. However, in the case of the method of checking the physical shooting environment, there is a disadvantage in that the user's accessibility is deteriorated because of the number of steps required by the user.

As described above, in some situations, such as when a third party attempts non-face-to-face authentication with an image of another person's ID recorded on a screen or printout, the actual ID image is obtained by classifying the image shooting environment with only one image input. A software or device that provides a function of determining is not provided. As a result, a waste of time and cost occurs, such as requiring an additional operation from the user.

A technical problem to be achieved through some embodiments of the present disclosure is to provide a method of receiving a target image and classifying a photographing environment of the corresponding image, and a computing system to which the method is applied.

Another technical problem to be achieved through some embodiments of the present disclosure is to determine the shooting environment of the target image by synthesizing data output from a plurality of artificial neural network models in order to provide high shooting environment classification accuracy for the target image. It is to provide a method and a computing system to which the method is applied.

Another technical problem to be achieved through some embodiments of the present disclosure is to provide a method of classifying a shooting environment of a target image according to characteristics such as texture and noise of the target image and a computing system to which the method is applied.

Another technical problem to be achieved through some embodiments of the present disclosure is to provide a method for preventing overfitting of learning an artificial neural network model that performs shooting environment classification for a target image.

Another technical problem to be achieved through some embodiments of the present disclosure is to provide a method of integrating a plurality of artificial neural network models receiving a plurality of segmented target images and a computing system to which the method is applied.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

In order to solve the above technical problem, a method for identifying whether a real-life captured image according to an embodiment of the present disclosure is configured to determine whether a first model learned to determine whether an entire image is a real-photographed image and whether a partial image is a real-photographed image Acquiring a learned second model; inputting data for a target image to the first model; obtaining first output data from the first model; image) into the second model, and acquiring second output data from the second model, and using both the first output data and the second output data, the target image is captured in real life. It may include outputting data on whether the image is an image.

In one embodiment, the first model and the second model include a first class indicating that the input image is a real-life shot image, a second class indicating that the input image is a display shot image, and a second class indicating that the input image is an output shot image. Probability values for each of the 3 classes can be output.

In one embodiment, the obtaining of the first output data from the first model may include reducing and resizing the original target image to a size of N x N (where N is a natural number greater than or equal to 2), and It may include inputting data into the first model.

In one embodiment, the obtaining of the second output data from the second model includes inputting data for the cropped image having a size of N x N, which is a partial region of the original target image, to the second model. can do.

In one embodiment, the step of inputting data for the cropped image having a size of NxN, which is a partial region of the original target image, to the second model, by dividing the original target image in a grid form to form a plurality of the cropped images. and inputting data of each of the plurality of cropped images to each of a plurality of detailed models of the second model.

In one embodiment, the extracting of the plurality of cropped images may include extracting an image of the region of interest by cropping the central portion of the original target image to a size of M X M, where M is the N * m (provided that, m is a natural number equal to or greater than 2) and extracting a plurality of cropped images by dividing the ROI image in a grid pattern.

In one embodiment, the plurality of cropped images may include a first cropped image, and obtaining second output data from the second model may include, for each of the plurality of cropped images, data of the cropped image. Iterating to determine the class of the cropped image by integrating the data output from each of the detailed models of the second model received, and synthesizing the class of each of the plurality of cropped images determined through the repeating step , obtaining the second output data.

In one embodiment, the number of detailed models of the second model may be smaller than the number of cropped images.

In one embodiment, the step of inputting the data of each of the plurality of cropped images to each of the plurality of detailed models of the second model may include randomly among the plurality of cropped images in each detailed model of the second model. and inputting data of a part of the cropped image sampled at the same time.

In one embodiment, the outputting of data on whether the target image is a real-life image includes the steps of assigning a first weight to the first output data and assigning a second weight to the second output data; The method may include summing the first output data and the second output data based on the assigned weight.

In one embodiment, the first weight and the second weight may be assigned based on brightness of the target image.

A real-life image identification system according to another embodiment of the present disclosure for solving the above technical problem may include one or more processors and a memory for storing one or more instructions, wherein the one or more processors, An operation of acquiring a first model learned to determine whether an entire image is a real-photographed image and a second model learned to determine whether a partial image is a real-photographed image by executing one or more stored instructions; An operation of inputting data to the first model and acquiring first output data from the first model; inputting data for a cropped image, which is a partial area of the target image, to the second model; An operation of obtaining second output data from a second model and an operation of outputting data on whether the target image is a real-life image by using both the first output data and the second output data may be performed. .

In one embodiment, the operation of obtaining the first output data from the first model may include an operation of reducing and resizing the original target image to a size of N x N (where N is a natural number greater than or equal to 2), and information on the resized target image. An operation of inputting data to the first model may be included.

In one embodiment, the operation of obtaining the second output data from the second model includes an operation of inputting data for the cropped image of size N x N, which is a partial region of the original target image, to the second model. can do.

In one embodiment, the operation of inputting data for the cropped image having a size of N x N, which is a partial region of the original target image, to the second model may include dividing the original target image in a grid form to form a plurality of cropped images. and inputting data of each of the plurality of cropped images to each of a plurality of detailed models of the second model.

In one embodiment, the operation of extracting the plurality of cropped images extracts an image of the region of interest by cropping the central portion of the original target image to a size of M X M, where M is the N * m (provided that, m is a natural number greater than or equal to 2) and extracting a plurality of cropped images by dividing the ROI image in a grid pattern.

In one embodiment, the plurality of cropped images may include a first cropped image, and the operation of obtaining second output data from the second model may include data of the cropped image for each of the plurality of cropped images. The operation of repeating determining the class of the cropped image by integrating the data output from each of the detailed models of the second model received and synthesizing the class of each of the plurality of cropped images determined through the repeating step , may include an operation of acquiring the second output data.

In one embodiment, the operation of outputting data on whether the target image is a real-life photographed image may include the operation of assigning a first weight to the first output data and assigning a second weight to the second output data; Based on the assigned weight, an operation of summing the first output data and the second output data may be included.

In one embodiment, the first weight and the second weight may be assigned based on brightness of the target image.

A computer program according to another embodiment of the present disclosure for solving the above technical problem is combined with a computing device and is trained to determine whether an entire image is a real-life image and whether a partial image is a real-life image. obtaining a second model learned to determine ?; inputting data of a target image into the first model and obtaining first output data from the first model; and providing data of a cropped image, which is a partial area of the target image, to the second model. input, obtaining second output data from the second model, and outputting data on whether the target image is a real-life photographed image by using both the first output data and the second output data. It may be stored in a computer-readable recording medium in order to be executed.

1 illustrates an exemplary environment to which a system for identifying whether a real-life photographed image according to an embodiment of the present disclosure may be applied.
FIG. 2 is a block diagram illustrating a system for identifying whether or not a real-life photographed image has been described with reference to FIG. 1 .
3 is a flowchart of a method for identifying whether or not a real-life photographed image is present according to another embodiment of the present disclosure.
4 is a diagram for illustratively explaining a learning process of a first real-life photographed image discrimination model for understanding some embodiments of the present disclosure.
5 and 6 are views for illustratively explaining a learning process of a second real-photographed image discrimination model for understanding some embodiments of the present disclosure.
7 is a diagram for exemplarily explaining a method of identifying whether or not a real-life image is captured by integrating a plurality of models for understanding some embodiments of the present disclosure.
FIG. 8 is a flowchart for explaining in detail the step of acquiring the first output data described with reference to FIG. 3 .
FIG. 9 is a flowchart for explaining in detail the step of obtaining second output data described with reference to FIG. 3 .
FIG. 10 is a diagram for exemplarily explaining a step of generating an output value for each cropped image from detailed models for understanding some embodiments of the present disclosure.
11 is a diagram for exemplarily explaining a method of randomly sampling a plurality of cropped images and inputting them to each detailed model of a second model in order to understand some embodiments of the present disclosure.
FIG. 12 is a flowchart for explaining in detail a data output step for determining whether the target image described with reference to FIG. 3 is a real-life photographed image.
13 is a hardware configuration diagram of a system for identifying whether or not a real-life photographed image is present according to another embodiment of the present disclosure.

Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Advantages and features of the present invention, and methods of achieving them, will become clear with reference to the detailed description of the following embodiments taken in conjunction with the accompanying drawings. However, the technical idea of the present invention is not limited to the following embodiments and can be implemented in various different forms, only the following embodiments complete the technical idea of the present invention, and in the technical field to which the present invention belongs It is provided to fully inform those skilled in the art of the scope of the present invention, and the technical spirit of the present invention is only defined by the scope of the claims.

In describing the present disclosure, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description will be omitted.

Hereinafter, several embodiments of the present disclosure are described with reference to the drawings.

1 is a diagram illustrating an environment to which a system for identifying a real-life captured image according to an embodiment of the present disclosure may be applied. As shown in FIG. 1, the system 100 for identifying whether a real-life captured image according to this embodiment interacts with the system 200 for identifying whether a real-life captured image or not and the user terminal 300 to implement a method for identifying whether a real-life captured image exists. can do.

Hereinafter, in order to help understanding of some embodiments of the present disclosure, a specific image to be identified as a real-photographed image of the real-photographed image identification system 100 will be described as a 'target image'.

In FIG. 1, the real-captured image identification system 100 and the real-captured image identification request system 200 are shown as separate systems, but in some embodiments, the real-captured image identification system 100 and the real-life The captured image identification request system 200 may be configured in a stand-alone manner in one system. In this case, it may be understood that the operation performed by the real-photographed image identification system 100 is performed by the real-photographed image identification request system 200 .

Hereinafter, each component shown in FIGS. 1 and 2 will be described in detail. In addition, in order to help understanding of some embodiments of the present disclosure, Detecting may mean identification of whether a target image is a real-life image.

The user terminal 300 may transmit the image to the system 200 for requesting identification of whether or not the actual photographed image is present. According to some embodiments of the present disclosure, the system 200 requesting identification of whether or not a real-life captured image may request the user terminal 300 to transmit an identification target image, but the user terminal 300 requests the system for identifying whether or not a real-life captured image ( 200) Alternatively, a user's input requesting identification of whether or not the actual captured image may be transmitted to the system 100 for identifying whether or not the actual captured image is present.

The user terminal 300 may receive information related to whether the target image is a real-life captured image or not from the real-life captured image identification request system 200 . According to some embodiments of the present disclosure, information related to whether the target image is a real-life image may be received from the system 100 for identifying whether or not the target image is a real-life image.

The real-captured image identification request system 200 may request the real-captured image identification system 100 to identify whether the target image received from the user terminal 300 is a real-captured image.

The actual photographed image identification request system 200 may request the user terminal 300 to transmit a target image. In addition, the real-photographed image identification request system 200 may transmit the target image received from the user terminal 300 to the real-photographed image identification system 100 .

The real shot image identification system 100 may receive a target image from the real shot image identification request system 200 . The received target image may be stored in the image storage unit 110 . In addition, the target image that has been identified as a real shot image may be stored in the learning data storage unit 130 .

The real-shot image identification system 100 may obtain a first model learned to determine whether a target image is a real-shot image and a second model learned to determine whether a partial image is a real-shot image. The operation of learning the first model and the second model may be performed by the identification model learning unit 120 shown in FIG. 2, and the data used for learning the first model and the second model is a learning data storage unit. It may be stored in (130).

Here, the partial image may be an image obtained by cropping a partial region of the target image. An operation of obtaining a partial image by cropping a partial region of the target image may be performed by the image pre-processing unit 150 . The partial image may be used interchangeably with terms such as a crop image. In addition, the step of learning the first model and the second model will be described later in detail.

Next, the real-shot image identification system 100 may input data of a target image to the first model and obtain first output data from the first model. Here, according to some embodiments of the present disclosure, the first output data is a probability value that the target image is an image of a real object, a probability value that the target image is an image of a screen, and the target image may include a probability value that is an image of a photographed output.

In some embodiments of the present disclosure, the screen may be understood as a display device such as a monitor, and the output may be understood as a paper output on which specific information is recorded by a device such as a printer, but is not limited thereto.

In addition, in some embodiments of the present disclosure, it can be understood that an operation performed by the system 100 for identifying whether or not a real-life photographed image is performed to determine whether an image of a specific person's ID card is a photographed image of a real-life ID card. However, it should be noted that this is only an example to help understanding of the present disclosure, and the purpose is not limited anywhere in a situation requiring identification of whether or not a real-life photographed image is required.

The real-shot image identification system 100 may input data about a cropped image of the target image to the second model and obtain second output data from the second model. The obtaining of the first output data and the second output data may be performed by the image identification unit 140 and will be described later.

The real-photographed image identification system 100 may output data on whether the target image is a real-photographed image by using both the acquired first and second output data. An operation of outputting data on whether the target image is a real-life image may be performed by the image identification unit 140 as shown in FIG. 2 . Also, according to some embodiments of the present disclosure, the real-shot image identification system 100 may evaluate whether or not the target image is a real-shot image by using data on whether the target image is a real-shot image.

So far, with reference to FIGS. 1 and 2 , the configuration and operation of the real-captured image identification system 100 and exemplary environments to which the real-captured image identification system 100 can be applied have been described.

Hereinafter, with reference to FIGS. 3 to 10 , a method for identifying a real-life captured image according to another embodiment of the present disclosure will be described in more detail. Hereinafter, steps described in some flowcharts may be understood to be performed by the system 100 for recognizing whether or not the actual photographed image is present unless otherwise specified.

In step S100 shown in FIG. 3, the system for determining whether a real-life image is captured acquires a first model learned to determine whether a target image is a real-photographed image and a second model learned to determine whether a partial image is a real-world captured image. can Hereinafter, a process of learning the first model and the second model according to some embodiments of the present disclosure will be described in detail with reference to FIGS. 4 to 7 .

According to some embodiments of the present disclosure, the full image learning model 40 shown in FIG. 4 may be an artificial neural network model trained to determine whether a target image is a real-life image. More preferably, it may be a Convolutional Neural Network (CNN) model trained to classify the environment in which the target image was captured.

Hereinafter, the full image learning model 40 will be referred to as a 'first model 40' to aid understanding of some embodiments of the present disclosure. The first model may receive data on a plurality of images of different subjects as training data. The learning data is, for example, in the case of the first model learned to determine whether or not an actual ID image is captured, an image of an actual ID card, an image of a paper printout on which the ID image is recorded, and a screen displaying the ID image. It may mean an image taken, but is not limited thereto.

In some embodiments of the present disclosure, data output by the first model trained to determine whether or not the ID image is actually captured may be a probability value corresponding to each environment category in which the ID image is captured. For example, the first model has a probability value that the input image is an actual ID card image, a probability value that is an image that captures a screen displaying the ID card image, and a probability value that is an image that captures a paper printout on which the ID image is recorded. All values can be printed.

In some embodiments of the present disclosure, the first model may receive a plurality of input images as training data in a resized state of NxN size. Here, the N value may be a natural number greater than or equal to 2 and may be a value smaller than the size value of the image input for learning the first model.

If the sizes of a plurality of images input as learning data are different, learning about the size of each of the plurality of images is required, and thus more learning data may be required to build a real-life image identification model. According to this embodiment, even if a plurality of images having arbitrary sizes are input as training data, the real-life image identification system adjusts the image to a square to build a real-life image identification model with higher accuracy with less training data. It can.

According to some embodiments of the present disclosure, the 'second model' is a plurality of artificial neural network models (artificial neural network models) learned to determine whether a cropped image, which is an image obtained by cropping a partial region of a target image shown in FIG. 5, is a real-life image ( 51, 52, 53) may be an integrated artificial neural network model. More preferably, it may be an artificial neural network model in which a plurality of CNN models learned to determine whether the cropped image is a real-life image are integrated. In addition, the number of the plurality of artificial neural network models may be a value predefined by a user or an initial value of the system.

According to some embodiments of the present disclosure, the cropped image may be at least one of images obtained by dividing a target image in a grid pattern. Referring to FIG. 6 , a method of dividing the target image into a lattice form will be clearly understood, and in the image attached to FIG. 6 , element 62 may correspond to one of a plurality of cropped images.

Hereinafter, to help understanding of some embodiments of the present disclosure, a plurality of artificial neural network models included in the second model are referred to as 'detailed models' or 'detailed models of the second model'.

However, the number of detailed models of the second model may be less than the number of cropped images.

The second model may receive data on a plurality of cropped images obtained by cropping a partial area of each of a plurality of images of different objects as training data. The learning data is, for example, in the case of the second model learned to determine whether or not an actual ID image is captured, a cropped image of an image of an actual ID card, a cropped image of an image of a paper printout on which the ID image is recorded, and It may mean a cropped image of an image taken of a screen displaying an ID image, but is not limited thereto.

In some embodiments of the present disclosure, as shown in FIG. 5 , the plurality of cropped images may be randomly sampled and input to a detailed model of the second model. A method of randomly sampling some of the plurality of cropped images and inputting them to the detailed model of the second model will be described later in detail.

In the case of the first model that identifies whether or not a real-life image is captured based on data on the entire image, accuracy may decrease when receiving an image that is different from an image encountered as training data. According to the present embodiment, features such as noise or texture of an image that are difficult to be analyzed by the first model may be analyzed through the second model that analyzes the cropped image of the input image in detail. Therefore, according to the present embodiment, a method for identifying whether or not a real-life photographed image has higher reliability and a computing system to which the method is applied may be provided.

According to some embodiments of the present disclosure, as shown in FIG. 7 , the real-life image identification system includes the full image learning model (first model) 40 and the split image learning model (second model) 50 It is possible to determine whether the input ID image is captured in real life by synthesizing the output data. The step of determining whether or not the real image was taken will be described later in detail.

So far, the step of learning the first model and the second model included in the real-life identification system has been described in detail. Hereinafter, the description will be continued with reference to FIG. 3 .

In step S200, the system for identifying whether or not a real-life shot image may input data on a target image to the obtained first model and obtain first output data from the first model. The obtaining of the first output data will be described with reference to FIG. 8 .

In step S210 shown in FIG. 8 , the system for identifying whether or not the actual captured image may resize the target image. Here, the resizing may mean adjusting the target image into an NxN square. Also, the N value may be a natural number greater than or equal to 2 and a value smaller than the size value of the target image.

If the first model is a model that has not been sufficiently trained to analyze an image having the size of the target image, a real-life image identification result may be incorrectly output. According to the present embodiment, by equally adjusting size values in all dimensions of the target image and inputting the values to the first model, an effect of increasing reliability of an image analysis result output by the first model may be achieved.

In step S220, the system for identifying whether or not the actual photographed image may obtain first output data by inputting the resized target image to a first model. The first output data may be a probability value corresponding to each environment category in which the target image is photographed. For example, when the first model is a model that classifies image capturing environments into first to third capturing environments, the first model analyzes the target image and sets a probability value of being captured in the first capturing environment to 77.2. %, a probability value of being photographed in the second photographing environment is 64.1%, and a probability value of being photographed in the third photographing environment is calculated as 42.0%, and a plurality of the calculated probability values may be output.

According to some embodiments of the present disclosure, the plurality of probability values output by the first model may be used in the step of outputting data on whether the target image is a real-shot image or not, but , which will be described later.

Hereinafter, the description continues with reference to FIG. 3 .

In step S300 of FIG. 3 , the real shot image identification system may input data about a cropped image, which is a part of the target image, to a second model, and obtain second output data from the second model. The obtaining of the second output data will be described in detail with reference to FIGS. 9 to 11 .

In step S310 of FIG. 9 , the system for identifying a real-life image may extract a plurality of cropped images from the target image. Here, the cropped image may refer to at least one of a plurality of images obtained by dividing the target image in a grid pattern. Referring to FIG. 10 , a method of extracting cropped images 910 and 920 by segmenting a target image in a lattice form will be clearly understood.

In some embodiments related to step S310, the system for identifying whether or not the actual captured image may extract a region of interest image from the target image. More specifically, when the target image is a rectangular image, since a plurality of cropped images having the same size cannot be extracted even if the target image is divided in the form of a grid, the real-life image identification system identifies the center of the target image as MxM. Can be cropped to size.

Here, M may be a value calculated by multiplying the size value of the target image by m. For example, when the target image is a rectangular image with a width of 1000 and a height of 1200, the real-life image identification system multiplies the height value of the target image by 0.75 and the width value of the target image by 0.8 to extract an 800x800 ROI image. It could be. More preferably, since the ROI image is an image obtained by cropping the central portion of the target image, the system for identifying whether or not a real-life image is taken is 200 pixels from the upper part, 200 pixels from the lower part, 100 pixels from the left side, and 100 pixels from the right side of the target image. 800x800 ROI images may be extracted by cropping each.

In some other embodiments related to step S310, the system for identifying whether the actual captured image may obtain a cropped image having a size of NxN by dividing the ROI image in a grid pattern. Here, N may be the same value as the size value of the resized target image input to the first model.

Next, in step S320, the system for identifying whether or not the actual captured image may input the extracted cropped image to a detailed model of the second model and generate an output value for each of a plurality of cropped images. A method of generating an output value for each of the plurality of cropped images will be described with reference to FIGS. 10 to 11 .

In addition, in some embodiments described with reference to FIGS. 10 to 11 below, the second model is illustrated as including first to third detailed models to aid understanding of the present disclosure. In addition, the second model outputs a probability value that the input image is a real-life image, a probability value that is an image of a screen displaying an ID image, and a probability value that is an image of a paper output on which an ID image is recorded. It should be noted that it is exemplified as an artificial neural network model, but is not limited thereto.

Referring to FIG. 10 , first to second cropped images 910 and 920 may be extracted in step S310. In some embodiments related to step S320, the first to third detailed models take the first cropped image 910 in real life in response to determining that the probability value of the first cropped image 910 is the highest. image can be determined.

In addition, in response to the determination that the number of detailed models for which the first cropped image 910 has been determined as a real-life image corresponds to a plurality of detailed models according to the determination results 911, 912, and 913 of the detailed models, It may be determined that the first cropped image 910 is a real image.

In some other embodiments related to step S320, the first detailed model determines that the second cropped image 920 is a real-life photographed image in response to determining that the probability value of the second cropped image 920 is the highest. The second and third detailed models may determine the second cropped image 920 as an output captured image in response to determining that the second cropped image 920 has the highest probability of being an output captured image. can

In addition, the actual photographed image identification system responds to a determination that the number of detailed models for which the second cropped image 920 has been determined as an output photographed image corresponds to a plurality according to the determination results 921, 922, and 923 of the detailed models. , it may be determined that the second cropped image 920 is a captured image of the printout.

In some other embodiments related to step S320, the system for identifying whether or not the real-life captured image determines the average sum of probability values that the first cropped image 910 is a real-life captured image output from the first to third detailed models is first The median probability value that the cropped image 910 is a real-life image, and the average of the sum of probability values that the first cropped image 910 is a screen-captured image is the medium probability value that the first cropped image 910 is a real-life image. , the average of the sum of probability values that the first cropped image 910 is an output captured image may be determined as an intermediate probability value that the first cropped image 910 is an output captured image.

The operation of determining the intermediate probability value may be performed on all of a plurality of cropped images.

In some other embodiments related to step S320, each of the first to third detailed models may receive data of a randomly sampled part of the cropped images among the plurality of cropped images.

For example, referring to FIG. 11 , among first to ninth cropped images extracted from the target image 90, a first detailed model includes first, second, and sixth cropped images 90-1, 90- 2, 90-6), the fourth, fifth, and ninth cropped images 90-4, 90-5, and 90-9 as the second detailed model, and the third, seventh, and ninth cropped images as the third detailed model. 8 cropped images (90-3, 90-7, 90-8) can be input. Here, the fact that each detailed model receives each cropped image is merely an example to help understanding of the present disclosure, and it will be understood that the number of cases of cropped images received by each detailed model is randomly designated. .

When the first to third detailed models repeatedly receive cropped images of the same region, a problem in that the accuracy of the actual image analysis result may decrease compared to the learned result may occur. According to this embodiment, the first to third detailed models receive similar but not identical data each time an image is analyzed, thereby preventing overfitting and improving the accuracy of an image analysis result.

Next, in step S330, the system for identifying whether or not the actual photographed image may generate second output data using data generated from the detailed models. Here, the second output data may include a probability value that a target image including a plurality of cropped images is a real image, a probability value that is a captured image, and a probability value that is a screen captured image.

In some embodiments related to step S330, the second model may determine the average of each of the intermediate probability values that the plurality of cropped images are output captured images determined in step S320 as the probability value that the target image is an output captured image.

In some other embodiments related to step S330, the second model may determine the average of each of the intermediate probability values that the plurality of cropped images determined in step S320 is a real-photographed image as the probability value that the target image is a real-photographed image.

In some other embodiments related to step S330, the second model may determine the average of each of the intermediate probability values that the plurality of cropped images are screen-captured images determined in step S320 as the probability value that the target image is a screen-captured image.

So far, the method for generating the second output data by the real-photographed image identification system has been described in detail. In addition, it should be noted that each step described so far may be used independently by the computing system, but a plurality of steps may be used together as needed.

The description continues with reference to FIG. 3 below.

In step S400, the real-life image identification system may output data on whether the target image is a real-photographed image by using both the first output data and the second output data. In more detail, it will be described with reference to FIG. 12 .

In step S410 shown in FIG. 12 , the system for identifying whether or not the actual photographed image is present may assign a weight to each of the first output data and the second output data.

In some embodiments related to step S410, in response to determining that the brightness or illuminance value of the input target image exceeds a maximum threshold value or is less than a minimum threshold value, the system for identifying whether or not the real-life image is captured provides second output data rather than first output data. Higher weights can be assigned.

The first model, which identifies whether or not a real-life image is captured based on data on the entire image, may have a problem in that accuracy decreases when an image that is different from the image encountered as training data is input. In general, the input image is bright. Alternatively, a case in which the illuminance is different from that of the training data image may correspond to this case. Therefore, according to the present embodiment, the system for identifying whether a real-life captured image can achieve an effect of being able to perform a method for identifying whether a real-life captured image has high accuracy even when receiving an image having an abnormal feature value.

Next, in step S420, the system for identifying whether or not the actual captured image may add the first output data and the second output data based on the assigned weight.

In some embodiments related to step S420, the first output data and the second output data may be calculated as follows.

단,

step,

a: first output data

b: second output data

: 제2 출력 데이터의 가중치

: Weight of the second output data

: 제1 출력 데이터의 가중치

: Weight of the first output data

Here, a and b of the above equation may be a probability value that the target image output by the first model is the first capturing environment image and a probability value that the target image output by the second model is the first capturing environment image, respectively. . In this case, the actual photographed image identification system may determine the result of the above equation as a final probability that the target image is the first photographed environment image.

In addition, a and b of the above equation may be a probability value that the target image output by the first model is a second capturing environment image and a probability value that the target image output by the second model is a second capturing environment image, respectively. . In this case, the actual photographed image identification system may determine the result of the above equation as a final probability that the target image is the second photographed environment image.

In addition, a and b of the above equation may be a probability value that the target image output by the first model is a third capturing environment image and a probability value that the target image output by the second model is a third capturing environment image, respectively. . In this case, the actual photographed image identification system may determine the result of the above equation as a final probability that the target image is the third photographed environment image.

Next, in step S430, the system for identifying whether the real-life captured image outputs data on whether the target image is a real-photographed image based on the sum of the first output data and the second output data.

In some embodiments related to step S430, the system for identifying whether or not the actual captured image may compare probability values that the target image is the first to third captured environment images calculated in step S420. For example, when it is assumed that the first photographing environment is a real photographing environment, the real photographed image identification system determines the target image to be real if the first photographing environment image probability value is higher than the second and third photographing environment image probability values. It is determined that it is a captured image, and data related thereto may be output.

In addition, in the above example, if at least one of the second or third capturing environment image probability values has a value higher than the first capturing environment image probability value, it is determined that the target image is not a real-life capturing image, and You can also output related data.

So far, a method for identifying whether a real-life photographed image has been described in detail according to another embodiment of the present disclosure. In addition, it should be noted that each step described so far may be used independently by the computing system, but a plurality of steps may be used together as needed.

13 is a hardware configuration diagram of a system for identifying whether or not a real-life photographed image is present according to some embodiments of the present disclosure. The actual photographed image identification system 1000 shown in FIG. 13 may refer to the real photographed image identification system 100 described with reference to FIG. 1 , for example. The system 1000 for identifying whether or not a real photographed image is a memory for loading a computer program 1500 executed by one or more processors 1100, a system bus 1600, a communication interface 1200, and the processor 1100 ( 1400) and a storage 1300 for storing the computer program 1500.

The processor 1100 controls the overall operation of each component of the system 1000 for identifying whether or not the actual photographed image is present. The processor 1100 may perform an operation for at least one application or program for executing a method/operation according to various embodiments of the present disclosure. Memory 1400 stores various data, commands and/or information. Memory 1400 may load one or more computer programs 1500 from storage 1300 to execute methods/operations according to various embodiments of the present disclosure. The bus 1600 provides a communication function between components of the system 1000 for identifying whether or not a real-life photographed image exists. The communication interface 1200 supports Internet communication of the system 1000 for identifying whether or not a real-life image is captured. Storage 1300 may non-temporarily store one or more computer programs 1500 . Computer program 1500 may include one or more instructions in which methods/operations according to various embodiments of the present disclosure may be implemented. When the computer program 1500 is loaded into the memory 1400, the processor 1100 may execute the one or more instructions to perform methods/operations according to various embodiments of the present disclosure.

In some embodiments, referring to FIG. 13, the system 1000 for identifying whether or not a real-life captured image is configured using one or more physical servers included in a server farm based on a cloud technology such as a virtual machine. It can be.

The computer program 1500 includes an operation of obtaining a first model learned to determine whether an entire image is a real-life photographed image and a second model learned to determine whether a partial image is a real-world photographed image, and storing data on a target image as An operation of inputting data into a first model and obtaining first output data from the first model; inputting data for a cropped image, which is a partial region of the target image, into the second model; An instruction for performing operations of acquiring second output data from and outputting data on whether the target image is a real-life image by using both the first output data and the second output data may include

So far, various embodiments of the present disclosure and effects according to the embodiments have been described with reference to FIGS. 1 to 13 . Effects according to the technical spirit of the present disclosure are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

The technical idea of the present disclosure described so far may be implemented as computer readable code on a computer readable medium. The computer program recorded on the computer-readable recording medium may be transmitted to another computing device through a network such as the Internet, installed in the other computing device, and thus used in the other computing device.

Although actions are shown in a particular order in the drawings, it should not be understood that the actions must be performed in the specific order shown or in a sequential order, or that all shown actions must be performed to obtain a desired result. In certain circumstances, multitasking and parallel processing may be advantageous. Although the embodiments of the present disclosure have been described with reference to the accompanying drawings, those of ordinary skill in the art to which the present disclosure pertains may implement the present invention in other specific forms without changing the technical spirit or essential features. can understand that there is Therefore, the embodiments described above should be understood as illustrative in all respects and not limiting. The protection scope of the present invention should be construed by the following claims, and all technical ideas within the equivalent range should be construed as being included in the scope of the technical ideas defined by the present disclosure.

Claims (20)

As a method performed in a computing system,
obtaining a first model learned to determine whether an entire image is a real-photographed image and a second model learned to determine whether a partial image is a real-photographed image;
inputting data of a target image into the first model and obtaining first output data from the first model;
inputting data for a cropped image, which is a partial region of the target image, into the second model, and obtaining second output data from the second model; and
Using both the first output data and the second output data, outputting data on whether the target image is a real-life photographed image,
How to identify whether the real-life image is taken. According to claim 1,
The first model and the second model are respectively for a first class indicating that the input image is a real-life photographed image, a second class indicating that the input image is a display photographed image, and a third class indicating that the input image is an output photographed image. which outputs a probability value,
How to identify whether the real-life image is taken. According to claim 1,
Obtaining first output data from the first model,
reducing and resizing the original target image to a size of NxN (where N is a natural number greater than or equal to 2); and
Including the step of inputting data for the resized target image to the first model,
How to identify whether the real-life image is taken. According to claim 3,
Obtaining second output data from the second model,
Including the step of inputting data for the cropped image of size N x N, which is a partial region of the original target image, to the second model.
How to identify whether the real-life image is taken. According to claim 4,
In the step of inputting data for the cropped image of size N x N, which is a partial area of the original target image, to the second model,
extracting a plurality of the cropped images by dividing the original target image in a grid pattern; and
Including inputting data of each of the plurality of cropped images to each of a plurality of detailed models of the second model,
How to identify whether the real-life image is taken. According to claim 5,
The step of extracting the plurality of cropped images,
extracting a region-of-interest image by cropping the central portion of the original target image to an MXM size, where M is the N*m (where m is a natural number greater than or equal to 2); and
Extracting a plurality of the cropped images by dividing the ROI image in a grid form,
How to identify whether the real-life image is taken. According to claim 5,
The plurality of cropped images include a first cropped image,
Obtaining second output data from the second model,
For each of the plurality of cropped images, repeating determining a class of the cropped image by integrating data output from each of the detailed models of the second model that has received data of the cropped image; and
Comprising the step of obtaining the second output data by integrating the classes of each of the plurality of cropped images, determined through the repeating step,
How to identify whether the real-life image is taken. According to claim 4,
The number of detailed models of the second model is smaller than the number of cropped images,
How to identify whether the real-life image is taken. According to claim 8,
The step of inputting data of each of the plurality of cropped images to each of a plurality of detailed models of the second model,
Including inputting data of randomly sampled partial cropped images among the plurality of cropped images to each detailed model of the second model.
How to identify whether the real-life image is taken. According to claim 1,
In the step of outputting data on whether the target image is a real-life image,
assigning a first weight to the first output data and assigning a second weight to the second output data; and
Based on the assigned weight, summing the first output data and the 2C output data,
How to identify whether the real-life image is taken. According to claim 10,
The first weight and the second weight,
Which is given based on the brightness of the target image,
How to identify whether the real-life image is taken. one or more processors; and
a memory for storing one or more instructions;
The one or more processors,
By executing one or more of the stored instructions,
obtaining a first model learned to determine whether an entire image is a real-photographed image and a second model learned to determine whether a partial image is a real-photographed image;
inputting data of a target image into the first model and obtaining first output data from the first model;
inputting data for a cropped image, which is a partial region of the target image, to the second model, and obtaining second output data from the second model; and
Performing an operation of outputting data on whether the target image is a real-life photographed image by using both the first output data and the second output data,
A system for identifying real-life images. According to claim 12,
Obtaining first output data from the first model,
reducing and resizing the original target image to a size of NxN (where N is a natural number greater than or equal to 2); and
Including the operation of inputting data for the resized target image to the first model,
A system for identifying real-life images. According to claim 13,
Obtaining second output data from the second model,
Including an operation of inputting data for the cropped image of size N x N, which is a partial region of the original target image, to the second model,
A system for identifying real-life images. According to claim 14,
The operation of inputting data for the cropped image of size N x N, which is a partial area of the original target image, to the second model,
extracting a plurality of cropped images by dividing the original target image in a grid pattern; and
Including an operation of inputting data of each of the plurality of cropped images to each of a plurality of detailed models of the second model,
A system for identifying real-life images. According to claim 15,
The operation of extracting the plurality of cropped images,
extracting a region-of-interest image by cropping the central portion of the original target image to an MXM size, wherein M is the N*m (where m is a natural number greater than or equal to 2); and
Including an operation of extracting a plurality of the cropped images by dividing the ROI image in a grid form,
A system for identifying real-life images. According to claim 15,
The plurality of cropped images include a first cropped image,
Obtaining second output data from the second model,
repeating, for each of the plurality of cropped images, determining a class of the cropped image by integrating data output from each of the detailed models of the second model that has received data of the cropped image; and
Comprising an operation of obtaining the second output data by synthesizing each class of the plurality of cropped images determined through the repeating step,
A system for identifying real-life images. According to claim 12,
The operation of outputting data on whether the target image is a real-life photographed image,
assigning a first weight to the first output data and assigning a second weight to the second output data; and
Based on the assigned weight, adding the first output data and the second output data.
A system for identifying real-life images. According to claim 18,
The first weight and the second weight,
Which is given based on the brightness of the target image,
A system for identifying real-life images. Combined with a computing device,
obtaining a first model learned to determine whether an entire image is a real-photographed image and a second model learned to determine whether a partial image is a real-photographed image;
inputting data of a target image into the first model and obtaining first output data from the first model;
inputting data for a cropped image, which is a partial region of the target image, into the second model, and obtaining second output data from the second model; and
Stored in a computer-readable recording medium to execute a step of outputting data on whether the target image is a real-life photographed image by using both the first output data and the second output data,
computer program.

Images